Multi-Level Time Series Forecasting With AI Attention Embeddings
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional time series forecasting techniques are time-intensive and suffer from significant memory complexity, making them inefficient for generating forecasts across multiple temporal granularities.
Innovation Solution
Implementing multi-level time series forecasting using artificial intelligence techniques, including determining entity-related and temporal features, creating embeddings, processing them with neural network-based attention, and generating forecasts across various granularities using one-shot learning and eigen-score-based attention to reduce time and memory complexity.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If conventional time series forecasting techniques are used, then forecasts can be generated, but the process is time-intensive and suffers from significant memory complexity
Solution Approach 1:
The patent segments the time series forecasting task into multiple temporal granularity levels (e.g., daily, weekly, monthly). Each level is processed independently through separate neural network models, allowing parallel computation and reducing the overall time complexity compared to processing all granularities sequentially with a single model.
Solution Approach 2:
The patent introduces a new dimension of temporal granularity hierarchy, transforming the forecasting problem from a single-level task into a multi-level hierarchical structure. This dimensional change enables the system to capture patterns at different time scales simultaneously, improving both efficiency and accuracy while reducing memory requirements through structured feature engineering.
2Reliability
If conventional time series forecasting techniques are used, then forecasts can be generated, but the system suffers from significant memory complexity
Solution Approach 1:
The patent applies local quality by designing specialized feature engineering and neural network architectures for each temporal granularity level. Each level uses tailored features and model configurations optimized for its specific time scale, rather than using a uniform approach across all granularities. This reduces memory complexity by avoiding the need to store and process all features at all levels simultaneously.
Solution Approach 2:
The patent implements dynamic feature selection and model configuration that adapts to different temporal granularities. The system dynamically adjusts which features are processed and how models are structured based on the specific granularity level being forecasted, optimizing memory usage while maintaining forecasting accuracy across multiple time scales.
Data Source
AI summary
Methods, apparatus, and processor-readable storage media for multi-level time series forecasting using artificial intelligence techniques are provided herein. An example computer-implemented method includes determining entity-related features and temporal features from at least a portion of one or more sets of time series data pertaining to at least one entity; creating multiple embeddings by encoding at least a portion of the entity-related features and at least a portion of the temporal features using at least one artificial intelligence-based embedding technique; processing the multiple embeddings using at least one neural network-based attention technique; generating one or more data forecasts across one or more temporal granularity levels by processing at least a portion of results from the processing of the multiple embeddings using at least one artificial intelligence-based categorization technique; and performing one or more automated actions based at least in part on the one or more data forecasts.